Contamination-safe synthetic Thai exam reasoning seed for a possible v26 fallback.
Rows: 400
Format: SFT chat JSONL with id, source, category, task_type, answer, verifier, metadata, messages
Options: exactly four Thai labels, (ก), (ข), (ค), (ง)
Answer balance: 100 examples for each label
Coverage: Thai reading inference, rhetoric/tone, grammar/semantics, Thai culture/history/civics/social studies, science in Thai, practical… See the full description on the dataset page:
https://huggingface.co/datasets/Jnx03/kanitakorn-thaiexam-v26-broad-reasoning-agent2-20260614.
